On November 18, 2026, Google introduced the third model in the Gemini 4 family — the Gemini 4 Neon neural network, which, according to the developer, provides "cutting-edge reasoning capabilities for complex, high-intensity workflows, doing so faster and at a lower cost." The model occupies an intermediate position between the previously released Gemini 4 Hydrogen, focused on fast execution of everyday tasks, and the flagship Gemini 4 Carbon, designed for the most complex and long-term projects. Neon is designed to combine reasoning depth with operational efficiency, filling the niche of tasks where simply generating an answer is not enough: the model needs to analyze context, use various tools, and execute a sequence of actions.

Performance and Benchmarks

According to Google, Gemini 4 Neon delivers performance on par with Carbon, and in a number of tasks surpasses it, while operating at Flash speed and requiring fewer computational resources. In the independent Artificial Analysis Intelligence Index, the model took third place with a score of 61 points at the maximum xhigh thinking level, trailing only larger and slower models. For context, OpenAI's GPT-6.2 Astra, released on the same day, took second place in the same ranking. In the Vals Index, Neon ranked second with 70.9%. The model also topped specialized tests such as Legal Research Bench and Tax Agent Bench, as well as Zapier's AutomationBench and Vending-Bench 2, where it scored $16,843 in simulated long-term business management.

Text, File, and Enterprise Data Processing

The developer emphasizes that Neon works better with texts than its predecessors, delivering clear and well-structured answers, and more accurately analyzes user-uploaded files — spreadsheets, PDF documents, and photos. The model can process voluminous materials, turning them into concise summaries and reports, and extract additional information from other Google ecosystem applications, including Gmail and Docs. When conducting research, Neon accesses internet sources to provide up-to-date information — for example, when comparing prices or preparing analytical materials with the latest market data. The model's context window is 1 million tokens, and four thinking levels allow the user to choose a balance between analysis depth and resource consumption.

Programming and Agentic Tasks

In programming tasks, Neon also shows high results: it took first place in the Artificial Analysis Coding Agent Index with 58.6 points, and scored 65.2% in Terminal-Bench 4.1. Google specifically notes the model's ability to debug and refactor code across multiple files and repositories simultaneously. In the Gemini app, Neon now powers the "Tasks" mode — a feature where the model independently performs multi-step work. Google recently made this mode available for free in six countries, expanding the accessibility of agentic scenarios to a wide audience.

Cost and Availability

The developer highlights cost-efficiency as a key competitive advantage of Neon. At the standard high thinking level, the model is 32% cheaper than Carbon for comparable tasks. Third-party tests show that its usage costs are 50–80% lower than competing advanced models. API pricing is set at $3 per million input tokens and $15 per million output tokens; cached input tokens cost $0.15 per million. The model is already available via the Gemini API, Google AI Studio, Android Studio, and Gemini for business. In the Gemini app and Google Antigravity, it can be used by subscribers of the paid AI Pro and AI Ultra tiers. There is no free access to the model. Officially, Gemini services do not operate in Russia: the company does not include the country in the list of regions where the Gemini app and Gemini API are available, and the Google AI Pro subscription cannot be paid for using Russian payment instruments.

Specialized Cybersecurity Version

In parallel with the main model, Google introduced Gemini 4 Neon for cyber, a version designed for cybersecurity tasks. It helps identify vulnerabilities, verify exploitability, and fix code errors. In the AutoBaxBench test, the model scored 76%, and in SEC-bench Pro — 90.5%. Access to this version will begin rolling out to Fairwind program participants in the coming days. Thus, Gemini 4 Neon covers a wide range of scenarios — from enterprise analytics and programming to specialized work with cybersecurity tasks — while maintaining price attractiveness compared to competitors' flagship solutions.